A capacitor polymer film thickness quality detection system and method

Abstract

The application belongs to the technical field of automatic quality control, and relates to a capacitor polymer film thickness quality detection system and method.The system comprises: an optical anomaly real-time detection module, which generates a defect event judgment instruction; a defect feature coding and positioning module, which encapsulates a hard logic form coding word and a space-time stamp to generate a defect space-time feature data packet; a physical driving parameter synthesis module, which extracts frequency domain coefficient mapping in the hard logic form coding word to synthesize customized physical driving parameters; an interception trigger signal generation module, which compares an absolute delay execution threshold value with a real-time motion main shaft pulse width count value to generate an interception trigger signal; an in-situ three-dimensional morphology acquisition module, which compiles physical microscopic three-dimensional morphology data; and a cross-modal fusion and traceability diagnosis module, which compares and judges an output terminal state defect traceability warning.The application solves the problem of dynamically adjusting the recognition standard of defect features to effectively control the misjudgment rate of different defect types when process parameter fluctuation leads to surface quality change.

Description

Technical Field

[0001] This invention belongs to the technical field of automated quality control, and relates to a capacitor polymer film thickness quality detection system and method. Background Technology

[0002] Vacuum coating technology is widely used in modern production, and the surface quality of the thin film directly affects the final performance of the product. In conventional production control processes, quality inspection typically relies on manual observation or basic image processing techniques for evaluation. During the coating operation, the manifestation of surface defects dynamically changes with fluctuations in process conditions. Traditional inspection methods have limitations in adapting to the diversity and dynamic changes in defect morphology, and are prone to deviations in inspection results when dealing with complex changes in the production environment.

[0003] Chinese patent application 202510977832.7 describes an intelligent monitoring method for vacuum coating quality based on artificial intelligence. This type of prior art mainly acquires basic process parameters through sensors in the coating equipment and combines them with visual image data to monitor surface quality. This technical solution typically relies on pre-set fixed rules to extract defect features, and then classifies and judges abnormal morphologies in the image based on static internal standards.

[0004] However, the aforementioned existing technologies still have room for improvement in practical applications. Due to the instability of process parameters during the coating process, the distribution density and directionality of defects exhibit significant non-uniform variations. The fixed identification rules in existing technologies are insufficient to accurately capture these subtle dynamic differences. When the monitoring system identifies different defect types, the lack of a mechanism to flexibly adjust judgment criteria according to real-time environmental changes easily leads to missed detections or false alarms for complex defect morphologies. This results in a high misjudgment rate under complex and fluctuating process conditions, making it difficult to achieve high-precision quality control in actual production.

[0005] Therefore, the technical problem that urgently needs to be solved in this field is how to dynamically adjust the identification criteria of defect features to effectively control the misjudgment rate of different defect types when the surface quality changes due to fluctuations in process parameters. Summary of the Invention

[0006] In a first aspect, the present invention provides a capacitor polymer film thickness quality detection system, comprising the following modules: To address the aforementioned problems, this invention provides a capacitor polymer film thickness quality detection system.

[0007] A capacitor polymer film thickness quality detection system, comprising: The real-time optical anomaly detection module acquires the internal refractive micro-transmission light flow through the polymer film and converts it into a continuous grayscale sequence. It calculates the total value of the light intensity gradient energy corresponding to the continuous grayscale sequence and generates a defect event judgment command when the total value of the light intensity gradient energy exceeds the preset energy threshold boundary. The defect feature encoding and localization module responds to defect event judgment instructions by cutting out solidified abnormal image frames from a continuous grayscale sequence, performing frequency domain mapping and splicing on the solidified abnormal image frames to generate hard logic morphology encoding words, combining preset high-precision pulse count values ​​and global timestamps to construct a spatiotemporal stamp, and encapsulating the hard logic morphology encoding words and spatiotemporal stamps to generate a defect spatiotemporal feature data packet. The physical driving parameter synthesis module unpacks the defect spatiotemporal feature data packet, separates the hard logic morphology code word and spatiotemporal stamp, and extracts the frequency domain coefficients in the hard logic morphology code word to synthesize customized physical driving parameters. The intercept trigger signal generation module acquires the real-time motion spindle pulse width count value, extracts the high-precision pulse count value inside the time stamp and superimposes it with the preset compensation offset pulse constant to generate an absolute delay execution threshold, and compares the absolute delay execution threshold with the real-time motion spindle pulse width count value to generate an intercept trigger signal. The in-situ three-dimensional topography acquisition module responds to the interception trigger signal and controls the probe to perform in-situ scanning detection according to the customized physical driving parameters. It receives the reflection spectrum decoding to calculate the absolute thickness difference information and compiles it into physical microscopic three-dimensional topography data. The cross-modal fusion and source tracing diagnosis module extracts the structural features of hard logic morphological encoded words to establish fingerprint database node data. It combines the fingerprint database node data with physical microscopic three-dimensional morphology data to calculate the optical microscopic thickness compatibility correlation coefficient value. The optical microscopic thickness compatibility correlation coefficient value is substituted into the preset offline production parameter failure model prior database for comparison and judgment, and the final state defect source tracing warning is output.

[0008] A further aspect of the present invention involves obtaining the internal refractive micro-transmission light flow through a polymer film and converting it into a continuous grayscale sequence, comprising the following steps: The driving schlieren illumination optical array projects a pre-defined structured illumination field onto the traveling polymer film; Simultaneously activate a high-speed linear array camera to capture the internally refracted microscopic transmitted light flow that is deflected through the polymer film; The spatial variation modes of the internal refractive micro-transmitted light flow are converted into digital light intensity amplitudes, and a continuous grayscale sequence reflecting the physical parameters of the first derivative of the polymer film thickness is constructed using the digital light intensity amplitudes.

[0009] A further aspect of the present invention involves generating a defect event determination instruction when the total energy value of the light intensity gradient exceeds a preset energy threshold boundary, comprising the following steps: The continuous grayscale sequence is scanned one by one using a sliding detection window with a pre-set arrangement pattern size, and the analysis pixel matrix space is extracted and separated. Calculate the approximate gray-level difference gradient of all pixels across the horizontal and vertical ends within the sliding detection window, calculate the square of each approximate gray-level difference gradient individually, and sum them up across the entire sliding detection window to obtain a continuous and smooth total value of light intensity gradient energy. The total value of light intensity gradient energy obtained in real time is continuously sent to the hardware digital comparator to match and check the preset energy threshold boundary. When the total value of light intensity gradient energy is detected to be greater than the energy threshold boundary, the pin feedback outputs a level flip signal to form and transmit the defect event judgment instruction.

[0010] A further aspect of this invention involves performing frequency domain mapping and splicing on solidified abnormal image frames to generate hard logic morphological codes, including the following steps: The solidified abnormal image frame is imported into a proprietary two-dimensional discrete cosine transform calculation core array to perform a two-dimensional spatial to frequency domain transformation, generating a transform frequency domain matrix that is laid out by energy coefficients of different frequency bands. Extract the set of low-frequency distribution coefficients that represent the width of the macroscopic geometric distribution of suspected morphological defects inside the transform frequency domain matrix, and simultaneously extract the set of high-frequency edge coefficients that express the sharpness attribute of the abrupt change of the contour edge in the high-frequency distribution region at the very end of the transform frequency domain matrix. A digital fixed-point alignment bit-width calling logic cross-shift dynamic calculation combination splicing merge low-frequency distribution coefficient set and high-frequency edge coefficient set is established, integrated, compressed and packaged into a binary mask block of fixed encapsulation length and distributed externally as a hard logic morphological encoding word.

[0011] A further aspect of this invention involves extracting frequency domain coefficients from hard logic morphological encoded words and mapping them to synthesize customized physical driving parameters, including the following steps: The low-frequency distribution coefficient set, which records the macroscopic surface expansion contour features of defects and is encapsulated inside the hard logic morphological encoded word, is extracted and transformed by linear translation, superposition, and scaling, and then assigned to the reference bias voltage reference value of the galvanometer digital-analog core power supply simulation reference. Extract the high-frequency edge coefficient set that represents the sharp intensity of micro-abrupt changes at the boundary of the quantization expression within the hard logic morphology encoded word. Normalize the high-frequency edge coefficient set to obtain the average absolute floating-point constant and perform inverse linear scaling to convert it into the high-frequency division ratio constant that the internal clock flow of the driving system depends on. The high-frequency division ratio constant value is inversely proportional to the coefficient value in the high-frequency edge coefficient set. By establishing a mathematical hierarchy-related polynomial superposition and overall allocation program, the reference bias voltage base value is allocated and limited to the maximum coverage limit of the main sampling window displacement. The high-frequency division ratio constant is changed to the physical position update refresh frequency beat parameter. The integrated mixed configuration guides the subsequent probe physical follow-up execution of exclusive customized physical drive parameters.

[0012] A further aspect of the present invention involves comparing the absolute delay execution threshold with the real-time motion spindle pulse width count value to generate an interception trigger signal, comprising the following steps: The high-precision pulse count value hidden within the spatial stamp payload data is used to characterize the origin position at the first appearance of the optical phase boundary of the marker. The internal high-speed register area is called to read the difference in physical distance between the proportionally magnified optical detection trigger center and the physical scanning execution station. The compensation offset pulse constant is calculated and entered by the measurement. The hardware direct adder synthesizer is used to combine and enhance the compensation offset pulse constant to the top of the high-precision pulse count value to obtain the expected absolute delay execution threshold. A closed-loop acquisition channel for pulse sequence of digital level of photoelectric drive of large rotating shaft in production is constructed to stably intercept and capture the latest changes in the real-time motion spindle pulse width count value. The zero-wait direct combination gate array is used to seamlessly connect and calibrate the absolute delay execution threshold with the real-time motion spindle pulse width count value. When the value is in the nanosecond-level clock slot corresponding to the matching bit, the rising edge of the level is captured by hardware edge detection logic and the interception trigger signal formed by the drive signal jump is released.

[0013] A further aspect of the present invention, responding to an interception trigger signal and controlling the probe to perform in-situ scanning detection according to customized physical drive parameters, includes the following steps: The power switching transistor amplifier circuit matrix network is awakened and activated by the power jump event that occurs at the front end of the interception trigger signal, which is driven by the bottom execution element that is waiting to be taken over. The customized physical drive parameters are unpacked and reassembled to generate a set of analog waveform electrical signals with smooth high and low undulations and simulated control trajectories. These signals are then streamed into the miniature two-dimensional galvanometer displacement servo electromagnetic coil that supports the rocker arm and changes its direction. By injecting a group of analog waveform electrical signals into the displacement servo electromagnetic coil of the miniature two-dimensional galvanometer, an asymmetric electromagnetic deflection force field with uneven distribution and rapid jumps is generated. This force commands the miniature two-dimensional galvanometer actuator to drive the confocal dispersion sensor probe mounted on the side of the miniature two-dimensional galvanometer actuator to repeatedly sweep across the set area following an irregular Lissajous figure scanning trajectory generated by the real-time non-periodic changes of customized physical driving parameters. This achieves a comprehensive and full-range inspection of the source of the anomaly, completing the required in-situ scanning detection.

[0014] A further aspect of this invention involves compiling physical microscopic three-dimensional morphological data, including the following steps: In conjunction with in-situ scanning detection, the confocal dispersive sensor probe is simultaneously activated to continuously acquire high-frequency dual-band abnormal spectral feedback data streams of polymer films that are irradiated through the membrane and subsequently reflected and peeled off by the internal structure of the polymer film. The combination of reflection wavelength characteristic values ​​is extracted from the dual-band abnormal spectral feedback data stream, which consists of the first peak center wavelength corresponding to the reflection of the upper surface of the polymer film and the second peak center wavelength corresponding to the reflection of the lower surface. The formula conversion of the offline calibration reference constant coefficient is called, and the absolute difference between the center wavelength of the first peak and the center wavelength of the second peak is compared and calculated. In this way, the absolute thickness difference information of a single detection micro point is decoded. All the absolute thickness difference information captured and collected in the point cloud spatial convergence mapping and integrated coordinate tracking grid are interwoven to construct a three-dimensional physical microscopic morphology data.

[0015] A further aspect of this invention involves calculating the optical microscopic thickness compatibility correlation coefficient by combining fingerprint database node data with physical microscopic three-dimensional morphology data, including the following steps: The hard logic morphology encoding is passed into the multi-layer feedforward convolutional neural network processing layer to activate and iterate, deeply mining the abnormal optical flow abrupt change phenomenon characteristics hidden in the gaps of high-dimensional structure. The fusion processing obtains the relative grayscale mapping fingerprint database node data that reflects the variation phenomenon generated by the damage of local microstructure. The optical morphology relative grayscale mapping fingerprint database node data is internally segmented and retained to specifically point to the texture extreme value peak vector that reflects the high density distribution of frequency domain energy and the steep change of tilt angle configuration. At the same time, the physical micro three-dimensional morphology data is extracted and quantified to reflect the spatial thickness anomaly benchmark comparison reference vector composed of the difference of the maximum steep upward gradient of the undulating height surface. The calibration and correction of the texture extreme peak vector and the comparison of the spatial thickness anomaly reference vector are forced to be registered and mapped within the same virtual multifaceted analytical coordinate system. The evolution of the two vectors is observed, and their pointing angles are aligned and synchronized. The deviation of the trajectory changes is measured by algebraic embedding of Pearson coefficients. The measurement model is used to calculate the mapping, quantitatively presenting the density of the spatial distribution variation and the level of linear fitting overlap compatibility between the two vectors. The output is an optical microscopic thickness compatibility correlation coefficient value that characterizes the internal substantial consistency confidence intensity.

[0016] Secondly, the present invention provides a method for detecting the thickness quality of a capacitor polymer film, comprising the following steps: S1. Obtain the internal refractive micro-transmitted light flow through the polymer film and convert it into a continuous grayscale sequence. Calculate the total value of the light intensity gradient energy corresponding to the continuous grayscale sequence. When the total value of the light intensity gradient energy exceeds the preset energy threshold boundary, generate a defect event judgment command. S2. The defect event judgment instruction cuts out the solidified abnormal image frame from the continuous grayscale sequence, performs frequency domain mapping and splicing on the solidified abnormal image frame to generate hard logic morphology code word, combines the preset high-precision pulse count value and global timestamp to construct a spatiotemporal stamp, and encapsulates the hard logic morphology code word and spatiotemporal stamp to generate defect spatiotemporal feature data packet. S3. Unpack the defective spatiotemporal feature data packet to separate the hard logic morphology code word and spatiotemporal stamp, and extract the frequency domain coefficients in the hard logic morphology code word to map and synthesize customized physical driving parameters; S4. Obtain the real-time motion spindle pulse width count value, extract the high-precision pulse count value inside the time stamp and superimpose the preset compensation offset pulse constant to generate an absolute delay execution threshold, and compare the absolute delay execution threshold with the real-time motion spindle pulse width count value to generate an interception trigger signal. S5 responds to the interception trigger signal and controls the probe to perform in-situ scanning detection according to the customized physical driving parameters. It receives the reflection spectrum decoding to calculate the absolute thickness difference information and compiles it into physical microscopic three-dimensional morphology data. S6. Extract the structural features of the hard logic morphological encoded words to establish fingerprint database node data. Combine the fingerprint database node data with the physical microscopic three-dimensional morphology data to calculate the optical microscopic thickness compatibility correlation coefficient value. Substitute the optical microscopic thickness compatibility correlation coefficient value into the preset offline production parameter failure model prior database for comparison and judgment, and output the final state defect tracing warning.

[0017] In summary, the present invention has the following beneficial technical effects: 1. By extracting frequency domain features from optical anomaly image frames and converting them into customized driving parameters for downstream physical scanning probes, the scanning strategy for physical measurements can be dynamically adjusted. Specifically, the system uses a set of low-frequency coefficients in hard logic morphology coding words to map the scanning amplitude of a two-dimensional galvanometer, and uses the sharpness of defect edges to set the scanning clock frequency. This feature enables the subsequent confocal spectroscopy measurement probe to perform wide-range low-frequency scanning on larger and smoother defects, and narrow-range high-frequency scanning on smaller and sharper defects. This mechanism concentrates limited physical measurement resources on the areas where optical anomalies have been detected, improving the ability to perform high-density, adaptive physical micromorphological characterization for defect regions of different morphologies.

[0018] 2. By calculating the correlation between optical characterization and physical morphology, this method provides more objective data support for the qualitative judgment of defects. It aligns the two-dimensional grayscale features of defects captured upstream with the three-dimensional physical microscopic morphology data collected downstream under a unified spatiotemporal reference. A correlation algorithm is then used to quantify the consistency between the optical feature vector and the physical contour gradient vector, thereby determining whether optical anomalies are accompanied by actual physical thickness changes. This feature can effectively distinguish between defects accompanied by internal physical deformation and optical artifacts that do not affect the physical performance of the product, thus reducing material waste caused by misjudgments.

[0019] 3. The defect feature vectors, detected in real time and verified through cross-modal fusion, are compared with a prior database of production parameter failure models constructed offline. By calculating the similarity between real-time defect features and typical defect clusters caused by process anomalies in a multi-dimensional feature space, the system can output warning information containing possible causes when the matching degree reaches a preset condition. This technical solution transforms thin film quality inspection from simple post-defect interception to online process monitoring and diagnosis, which helps technicians quickly locate and correct potential problems in the production process.

[0020] 4. For high-speed moving polymer films, a positioning and interception logic combining a spatiotemporal stamp and downstream encoder counts was designed. At the instant a defect is initially detected optically, the system synchronously latches the high-precision pulse count value of the production line spindle encoder to form a spatiotemporal stamp, and adds it to the pulse constant corresponding to a fixed distance between upstream and downstream stations to calculate the absolute delay execution threshold. Using a downstream hardwired digital comparator to continuously compare the real-time encoder value with this threshold, a trigger signal can be output at the nanosecond level when the target defect physically reaches the second measurement station. This pure hardware position tracking and triggering mechanism, which avoids software interrupt responses, ensures the positioning accuracy for secondary physical measurements of tiny areas on the high-speed moving film. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0022] Figure 1 This discloses a schematic diagram of the framework in the embodiments of this application.

[0023] Figure 2 This discloses a flowchart of an embodiment of this application.

[0024] Figure 3 The present application discloses the dual-band reflectance spectral polarization characteristic map of confocal dispersive sensing in the embodiments of this application. Detailed Implementation

[0025] The following is in conjunction with the appendix Figure 1 - Figure 3 A preferred description of the present invention is provided below.

[0026] See attached document Figure 1 - Figure 2 This invention proposes a capacitor polymer film thickness quality detection system, comprising the following modules: The real-time optical anomaly detection module acquires the internal refractive micro-transmission light flow through the polymer film and converts it into a continuous grayscale sequence. It calculates the total value of the light intensity gradient energy corresponding to the continuous grayscale sequence and generates a defect event judgment command when the total value of the light intensity gradient energy exceeds the preset energy threshold boundary. The defect feature encoding and localization module responds to defect event judgment instructions by cutting out solidified abnormal image frames from a continuous grayscale sequence, performing frequency domain mapping and splicing on the solidified abnormal image frames to generate hard logic morphology encoding words, combining preset high-precision pulse count values ​​and global timestamps to construct a spatiotemporal stamp, and encapsulating the hard logic morphology encoding words and spatiotemporal stamps to generate a defect spatiotemporal feature data packet. The physical driving parameter synthesis module unpacks the defect spatiotemporal feature data packet, separates the hard logic morphology code word and spatiotemporal stamp, and extracts the frequency domain coefficients in the hard logic morphology code word to synthesize customized physical driving parameters. The intercept trigger signal generation module acquires the real-time motion spindle pulse width count value, extracts the high-precision pulse count value inside the time stamp and superimposes it with the preset compensation offset pulse constant to generate an absolute delay execution threshold, and compares the absolute delay execution threshold with the real-time motion spindle pulse width count value to generate an intercept trigger signal. The in-situ three-dimensional topography acquisition module responds to the interception trigger signal and controls the probe to perform in-situ scanning detection according to the customized physical driving parameters. It receives the reflection spectrum decoding to calculate the absolute thickness difference information and compiles it into physical microscopic three-dimensional topography data. The cross-modal fusion and source tracing diagnosis module extracts the structural features of hard logic morphological encoded words to establish fingerprint database node data. It combines the fingerprint database node data with physical microscopic three-dimensional morphology data to calculate the optical microscopic thickness compatibility correlation coefficient value. The optical microscopic thickness compatibility correlation coefficient value is substituted into the preset offline production parameter failure model prior database for comparison and judgment, and the final state defect source tracing warning is output.

[0027] In one embodiment of the present invention, the real-time optical anomaly detection module is configured to perform the following steps: The continuous grayscale sequence is scanned one by one using a sliding detection window with a pre-set arrangement pattern size, and the analysis pixel matrix space is extracted and separated. Calculate the approximate gray-level difference gradient of all pixels across the horizontal and vertical ends within the sliding detection window, calculate the square of each approximate gray-level difference gradient individually, and sum them up across the entire sliding detection window to obtain a continuous and smooth total value of light intensity gradient energy. The total value of light intensity gradient energy obtained in real time is continuously sent to the hardware digital comparator to match and check the preset energy threshold boundary. When the total value of light intensity gradient energy is detected to be greater than the energy threshold boundary, the pin feedback outputs a level flip signal to form and transmit the defect event judgment instruction.

[0028] Specifically, the first step of the method of the present invention involves acquiring and processing optical information from the surface of a polymer film moving at high speed to generate preliminary instructions for defect determination. This step is executed by a data processing and control unit, such as a control board integrating a field-programmable gate array (FPGA) and an embedded processor. The data processing and control unit drives a linearly arranged schlieren illumination optical array covering the full width of the polymer film. A schlieren illumination optical array is an optical system used to visualize refractive index gradients in a transparent medium. In this embodiment, it consists of a row of high-brightness white LEDs, a collimating lens group, and a Foucault knife edge or graded neutral density filter located at the focal position. The array projects a pre-defined structured illumination field, such as highly uniform parallel light generated by the collimating lens group, onto the polymer film moving at a constant speed V on the production line.

[0029] When the structured illumination field penetrates the polymer film, the minute refractive index gradient caused by thickness inhomogeneity or material density differences within the film deflects the transmitted light. Synchronously with the illumination array, the data processing and control unit activates a high-speed linear scan camera, located on the opposite side of the polymer film, with its photosensitive element aligned with the illumination array. This high-speed linear scan camera is an industrial camera with a single or multiple rows of photosensitive pixels. Its line scanning frequency is set in the range of 50 kHz to 200 kHz to match the typical production line speed of 100 m / min to 500 m / min for polymer films. This value is set within this range to balance detection resolution and system processing bandwidth. For example, with a typical film speed of 300 m / min (5000 mm / s) and a line frequency of 100 kHz, the spatial sampling resolution of the film in the direction of motion can reach [value missing]. This resolution matches the microscopic physical size of typical gel points or impurity defects in capacitor polymer films, typically between 0.1 mm and 0.5 mm, ensuring that any tiny defects span at least 2 to 3 pixels, thus avoiding missed detections caused by Nyquist undersampling.

[0030] With this configuration, a high-speed linear scan camera captures the invisible internal refractive micro-transmitted light flow, which varies spatial intensity due to light deflection, at a line scan frequency matched to the movement speed of the polymer film. Subsequently, an image capture circuit deployed on the data processing and control unit converts the analog photon energy signal corresponding to the internal refractive micro-transmitted light flow captured by the high-speed linear scan camera's image sensor into multi-bit depth digital light intensity amplitudes in real time via a high-speed data interface such as Camera Link or CoaXPress. These amplitudes are organized into a two-dimensional data matrix, where one dimension corresponds to the width and position of the thin film, and the other dimension corresponds to time or displacement along the film's movement direction.

[0031] This process constructs a series of time-continuous one-dimensional arrays, where each element represents the positional light intensity—a complete continuous grayscale sequence reflecting the physical information of the first derivative of the film thickness. Specifically, the continuous grayscale sequence is a set of one-dimensional numerical arrays arranged in time order, each array containing, for example, 8192 pixels, each pixel having a grayscale depth of 12 bits, representing 4096 grayscale levels.

[0032] Inside the FPGA, a parallel processing logic unit performs real-time analysis on a continuous grayscale sequence. This logic unit uses a sliding detection window with a preset size of W pixels multiplied by H pixels to scan the entire sequence pixel by pixel. The sliding detection window is a logical analysis region, the size of which is... The setting is based on the typical defect size to be detected; for example, to capture tiny gel point defects, it can be set to... The pixel size is based on the fact that, at a resolution of 0.05 mm / pixel, the actual physical area corresponding to 16×16 pixels is 0.8 mm × 0.8 mm. This window size is slightly larger than the maximum projected area of ​​most typical thin film defects mentioned above. This ensures that the abnormal optical envelope of the isolated defect is completely wrapped in one go, without introducing too much normal background noise due to an excessively large window, which would lead to a decrease in the signal-to-noise ratio.

[0033] At each position the sliding detection window moves, this logic unit calculates the total light intensity gradient energy of all pixels within the window in real time. Specifically, for each pixel within the window, it calculates the approximate gray-level gradient in both the horizontal and vertical directions, calculates the square of the gradient magnitude, and then sums these values ​​for all pixels within the entire window to obtain the total light intensity gradient energy. To further illustrate, in the calculation of the total accumulated light intensity gradient energy within the aforementioned sliding detection window, the total light intensity gradient energy... It is based on continuous grayscale sequences The grayscale value distribution within the window is calculated. The specific calculation formula is:

[0034] In the formula, The coordinates within the sliding detection window are The grayscale value of the pixel; and These represent the gradient components at the point in the horizontal and vertical directions, respectively, and can be approximated using first-order differences, for example... as well as ,in, and These represent the grayscale values ​​of the pixels preceding the current point horizontally and vertically, respectively. The summation operation iterates through all pixels within the sliding detection window, which has a size of [value missing]. The physical meaning of this formula lies in quantifying the complexity of image texture or the drastic degree of light intensity change within the detection window.

[0035] Finally, the hard-wired digital comparator integrated within the FPGA compares the real-time calculated total energy of the light intensity gradient with a configurable energy threshold boundary. Continuous comparisons are performed. It should be noted that this energy threshold boundary... The setting is based on the average and standard deviation of the background gradient energy obtained from statistical analysis of a large number of qualified thin film samples. It is usually set as the mean plus... Multiple standard deviations, of which This is an adjustable coefficient, ranging from 3 to 6. Assuming the application scenario is a production environment with low background noise, to obtain higher detection sensitivity, the following can be used: Set to 3.5, this coefficient The value of is based on the Gaussian normal distribution anomaly detection theory in statistics. The light intensity gradient fluctuation in the background region of a qualified thin film exhibits Gaussian white noise. The coefficient is then... Setting it to 3 or higher means that only abrupt changes that deviate from the background mean by more than 3 standard deviations will be considered defects. Mathematically, this guarantees that the theoretical false alarm rate of the system under pure background noise is less than 0.27%. For high-precision capacitor film detection, a value of 3.5 can achieve a balance between high sensitivity and anti-interference capability against small disturbances.

[0036] The dynamic energy threshold is dynamic because it can be adjusted online via a host computer software interface according to different batches or specifications of thin film products. Once the total energy value of the light intensity gradient is detected to exceed the energy threshold boundary within a certain clock cycle, the comparator immediately outputs a single-bit high-level logic signal. This signal is the defect event determination instruction and is sent to the subsequent processing unit as a hardware interrupt or trigger signal. The defect event determination instruction is a single-cycle TTL level pulse signal with a high-level duration not exceeding 100 ns to ensure the triggering accuracy of subsequent processing flows.

[0037] For example, suppose a polymer film on a production line travels at a speed of 300 m / min, using a high-speed linear scan camera with a line frequency of 100 kHz. At a certain moment, an internal gel point defect with a diameter of approximately 100 μm enters the camera's field of view. This defect causes a sharp change in the internal refractive micro-transmitted optical flow in its corresponding region. Reflected in a continuous grayscale sequence, the grayscale value of a local area rapidly drops from the background average of 1800 to 500 and then rises back to near 1800.

[0038] Size is The sliding detection window of the pixels is scanning the sequence in real time. When the window covers a normal area, the grayscale value changes. Its gradient component remains basically constant at around 1800. and Approaching 0, the calculated total energy value of the light intensity gradient is... To stabilize at a low value below 5000. When the sliding detection window moves to the area covering the gel point defect, due to the drastic jump in grayscale value, for example from 1800 to 500, the gradient difference of one pixel can reach 1300, with a square of 1690000. After accumulation within the window, the total value of the light intensity gradient energy is calculated. It surged to 9,500,000.

[0039] In this embodiment, the energy threshold boundary The cleaning sample for this type of film has been pre-set to 1,000,000. Due to the calculated... The value is 9,500,000, which is greater than the preset threshold. The digital comparator logic within the FPGA is immediately triggered, and the level of its output pin flips from logic 0 to logic 1, generating a valid defect event determination instruction. This instruction is then sent to the trigger input of the hardware module that will execute subsequent steps.

[0040] In one embodiment of the present invention, the defect feature encoding and localization module is used to perform the following steps: The solidified abnormal image frame is imported into a proprietary two-dimensional discrete cosine transform calculation core array to perform a two-dimensional spatial to frequency domain transformation, generating a transform frequency domain matrix that is laid out by energy coefficients of different frequency bands. Extract the set of low-frequency distribution coefficients that represent the width of the macroscopic geometric distribution of suspected morphological defects inside the transform frequency domain matrix, and simultaneously extract the set of high-frequency edge coefficients that express the sharpness attribute of the abrupt change of the contour edge in the high-frequency distribution region at the very end of the transform frequency domain matrix. A set of low-frequency distribution coefficients and high-frequency edge coefficients are established by dynamically calculating, combining, splicing, and merging the digital fixed-point alignment bit-width calling logic across displacements. These are then integrated, compressed, and packaged into a binary mask block of fixed encapsulation length for external distribution as a hard logic morphological encoding word.

[0041] Specifically, following the real-time optical anomaly detection module, when a defect event determination command is generated, the second implementation step of this method is triggered at the hardware level to perform multi-dimensional data space truncation and time-frequency conversion encoding on the continuous grayscale sequence. This process is implemented by an event response and data encoding unit, such as a logic area within the aforementioned FPGA, which responds to the rising edge level toggling signal of the defect event determination command. This level toggling signal acts as a synchronous latch enable signal, acting on a hardware logic gate array. This gate array controls the read / write pointers of a first-in-first-out FIFO or a dual-port RAM storing the continuous grayscale sequence.

[0042] Upon receiving the enable signal, the gate array freezes write operations to the memory and, centered on the address corresponding to the trigger moment, executes operations based on preset two-dimensional size parameters, such as... A pixel is extracted from a continuous grayscale data stream, forming a static two-dimensional data block. This data block contains an envelope composed of anomalous optical textures, creating a solidified anomalous image frame. Specifically, this solidified anomalous image frame is a two-dimensional digital matrix with a size of... Set as, for example There are two reasons for this. First, because the core algorithm of the two-dimensional discrete cosine transform is based on the Fast Fourier Transform architecture, the size is an integer power of 2, such as... Firstly, the hardware logic circuit's operational efficiency is improved, meeting the nanosecond-level timing requirements of online real-time detection. Secondly, by taking the 16×16 defect core area as the center and cutting out a 64×64 region, a sufficiently wide surrounding normal thin film area can be retained as a low-frequency background reference, which helps to generate a more contrasting low-frequency texture coefficient in the frequency domain space, ensuring that the defect core and its surrounding area detected by the optical anomaly real-time detection module are included.

[0043] After the solidified abnormal image frame is formed, it is loaded into the on-chip memory. Subsequently, the FPGA's internal processing pipeline injects the solidified abnormal image frame into the hardware-implemented two-dimensional discrete cosine transform (DCT) calculation core. The aforementioned two-dimensional discrete cosine transform is an orthogonal transform that converts a signal from the spatial domain to the frequency domain. This DCT core will... The spatial domain image matrix of pixels is transformed and mapped to the frequency domain to generate a two-dimensional DCT coefficient matrix of the same size. In this coefficient matrix, the coefficients in the upper left corner represent the low-frequency information of the image, reflecting the overall contour and gradually changing regions; while the coefficients in the lower right corner represent the high-frequency information of the image, reflecting abrupt changes in details such as edges and noise.

[0044] A coefficient selection and concatenation logic unit extracts two sets of coefficients from the DCT coefficient matrix. The first set is a fixed set of M low-frequency distributed coefficients, for example, selecting coefficients from the DCT coefficient matrix at coordinates... The DC component and its neighboring The set of communication components represents the geometric dimensions and basic morphology of the defects; the low-frequency distribution coefficient set is a subset of data characterizing the macroscopic features of the image, and its quantity... The value is usually set to 3 to 5. The basis for this setting is that the macroscopic energy of the image is mainly concentrated in the DC component in the upper left corner of the DCT matrix and the first few very low frequency AC components. Extracting the first 3 to 5 components can represent more than 90% of the macroscopic contour energy. Continuing to increase the number of extractions will result in a sharp decrease in the correction benefit of the galvanometer amplitude, which is sufficient to describe the approximate size and energy distribution of the defect.

[0045] The second group consists of a fixed set of K high-frequency edge coefficients, for example, those selected along a zigzag scanning path and located at the end of the coefficient matrix. The set of coefficients represents the sharpness of defect edges; the high-frequency edge coefficient set is a subset of data representing image detail features, and its quantity... The value is usually set to 4 to 8. The reason for this setting is that the extremely high frequency coefficients at the end of the zigzag scanning path are easily affected by random level noise. Extracting 4 to 8 coefficients and averaging them can effectively smooth out random noise, accurately extract the feature values ​​that represent the true sharpness of the defect edge, and at the same time, it does not occupy too much SPI bus communication bandwidth for quantifying the clarity of the defect edge.

[0046] This logic unit, according to a preset bit width (e.g., quantizing each coefficient to a 16-bit fixed-point number), concatenates the low-frequency distributed coefficient set with the high-frequency edge coefficient set into a hard logic morphological codeword through bit shifting and logical OR operations. This hard logic morphological codeword is a fixed-length binary data block, and its total bit width is determined by the number of selected coefficients and the bit width of each coefficient. bit.

[0047] Simultaneously, at the absolute trigger moment of the defect event determination command—that is, the absolute trigger moment when the solidified abnormal image frame is stripped and cut out—the system clock synchronization and external interface unit perform two parallel latching operations. First, they lock the current high-precision pulse count value sent from the production line spindle encoder to the LVDS interface via a high-speed differential signal; second, they lock the global timestamp generated by a high-precision clock source driven by a global GPS synchronization or crystal oscillator. These two values ​​are integrated to generate a unique spatiotemporal stamp.

[0048] Specifically, the spatiotemporal stamp is a composite data structure that combines spatial location and time information. The high-precision pulse count value provides the precise physical location along the direction of the film's movement, while the global timestamp provides the world-unified time of the event.

[0049] Finally, the data link layer protocol encapsulation unit, such as a hardware module that simplifies Ethernet MAC layer functions, merges and assembles the spatiotemporal stamp and hard logic morphology code according to a predefined data frame format. This process includes adding a start-of-frame delimiter, destination address, packet type identifier, and cyclic redundancy check (CRC) code, encapsulating and outputting a defect spatiotemporal characteristic data packet with a unique identifier. The defect spatiotemporal characteristic data packet is a data frame that conforms to a communication protocol, and its structure is designed to be efficiently parsed by downstream digital signal processors.

[0050] For example, continuing from the example in S1, when a defect event determination instruction is generated, the FPGA's internal logic responds immediately. Assume that at this moment, the high-precision pulse count sent by the production line spindle encoder is... The global timestamp of the global clock source is ms. These two are latched and combined to form a spacetime stamp.

[0051] Simultaneously, the system extracts a point from the continuous grayscale sequence, centered on the point with the largest energy gradient in S1. A fixed, aberration image frame. This image frame is fed into the DCT core. After transformation, it is assumed that the coefficient selection logic is configured to extract... Low-frequency coefficients and Several high-frequency coefficients. The DC component is extracted from the generated DCT coefficient matrix. And the first exchange component As a set of low-frequency distribution coefficients; and extract two high-frequency components. and As a set of high-frequency edge coefficients.

[0052] Each coefficient value is quantized as a 16-bit signed integer. The concatenation logic unit performs the following operations to generate a 64-bit hard logic morphological codeword. :Will Shift the value left by 48 bits, and Shift the value left by 32 bits, and... Shift the value left by 16 bits, and finally... The values ​​are then subjected to a logical OR operation. The data packet encapsulation unit encapsulates the frame header and hard logic morphology code words. and spacetime stamp The data is assembled together, and a CRC checksum is added to form a complete defect spatiotemporal characteristic data packet, which is then sent out through a gigabit Ethernet interface. The content of this data packet can be represented as: [Header|Data_Type|Content:{ ,( )}|CRC].

[0053] In one embodiment of the present invention, the physical drive parameter synthesis module is configured to perform the following steps: The low-frequency distribution coefficient set, which records the macroscopic surface expansion contour features of defects and is encapsulated inside the hard logic morphological encoded word, is extracted and transformed by linear translation, superposition, and scaling, and then assigned to the reference bias voltage reference value of the galvanometer digital-analog core power supply simulation reference. Extract the high-frequency edge coefficient set that represents the sharp intensity of micro-abrupt changes at the boundary of the quantization expression within the hard logic morphology encoded word. Normalize the high-frequency edge coefficient set to obtain the average absolute floating-point constant and perform inverse linear scaling to convert it into the high-frequency division ratio constant that the internal clock flow of the driving system depends on. The high-frequency division ratio constant value is inversely proportional to the coefficient value in the high-frequency edge coefficient set. By establishing a mathematical hierarchy-related polynomial superposition and overall allocation program, the reference bias voltage base value is allocated and limited to the maximum coverage limit of the main sampling window displacement. The high-frequency division ratio constant is changed to the physical position update refresh frequency beat parameter. The integrated mixed configuration guides the subsequent probe physical follow-up execution of exclusive customized physical drive parameters.

[0054] Specifically, after the aforementioned defect feature encoding and localization module outputs the defect spatiotemporal feature data packet, the third implementation step aims to deconstruct the data packet and synthesize the driving parameters used to control the downstream physical equipment. This step is implemented by a digital signal processor (DSP), such as a high-speed processor equipped with a floating-point unit and a multi-channel direct memory access (DMA) controller.

[0055] The Digital Signal Processor (DSP) executes low-level instruction-driven data reception assertion operations through its integrated physical layer interface, such as a Gigabit Ethernet MAC. This data reception assertion operation is a hardware feature of the DSP that allows the Network Interface Controller (NIC) to autonomously identify data packets and trigger processing flows without consuming CPU resources. This operation triggers a hardware interrupt when it detects a defective spatiotemporal characteristic data packet carrying a data type identifier destined for its own address. The DSP's interrupt service routine then starts, performing physical layer unpacking of the multi-byte data packet, including verifying the Cyclic Redundancy Check (CRC) code to ensure data integrity.

[0056] After verification, the DSP separates the data packet from the payload area in situ with a fixed byte offset according to the predefined data frame structure, loads the hard logic morphology code word representing the optical morphology of the defect into the internal cache, and stores the spatiotemporal stamp containing position and time information into the designated register group.

[0057] The DSP's arithmetic logic unit begins parallel processing of the hard logic morphological codeword. On one hand, it extracts and decomposes the low-frequency distribution coefficient set representing the defect geometry width from the hard logic morphological codeword, i.e., the first M coefficient values. For each coefficient value in this set, for example... and The DSP performs independent scaling and displacement operations on it, based on the following linear mapping relationship. The calculation is performed. The result is mapped and assigned to the reference bias voltage reference value of the digital-to-analog converter (DAC) in a downstream two-dimensional galvanometer drive controller.

[0058] To further illustrate, in the process of mapping the low-frequency distributed coefficient set, the reference bias voltage value is used. and Each is determined by its corresponding low-frequency coefficient value. and The calculation is as follows: ;

[0059] in, , and 、 These are the scaling and displacement coefficients corresponding to the X and Y axes, respectively, constants determined through system calibration. Their setting is based on linearly mapping the DCT coefficients within their statistical distribution range, for example -4096 to +4095, to the rated input voltage range of the 2D galvanometer driver controller (DAC), for example -5V to +5V, to achieve the maximum dynamic scanning range. Furthermore, in the voltage mapping formula... These are DCT coefficient values, with the scaling factor in V / count and the displacement factor in V. Therefore, the final... The unit is volts (V). This reference bias voltage is actually a set of analog DC voltages that set the peak values ​​of the X-axis and Y-axis scan signals for the downstream trajectory generation circuitry. This transforms abstract frequency domain coefficients into concrete analog voltage references that control the scan amplitude.

[0060] On the other hand, the DSP extracts the set of high-frequency edge coefficients representing the sharpness of defect edges from the hard logic morphological encoded word, namely the last K coefficient values. The DSP calculates the average amplitude of this coefficient set and converts it into a high-frequency division ratio constant used to set the scanning main frequency clock through another linear transformation. .

[0061] Specifically, when converting the high-frequency edge coefficient set, the high-frequency division ratio constant... From this set of coefficients The average absolute value of the coefficients is calculated, and the formula is as follows:

[0062] In the formula for the frequency division proportionality constant The first high-frequency edge coefficient set represents the... Each coefficient value is dimensionless. and All are dimensionless constants. Through this inverse proportional mapping, the sharper the defect edge, i.e., the larger the high-frequency coefficient, the smaller the calculated frequency division constant, thus generating a higher scanning frequency, and therefore the final... This is also an integer value, conforming to the parameter requirements of the frequency divider. It should be understood that this high-frequency division ratio constant is an integer value, loaded into the register of the DSP or an external clock management unit, used to divide a high-frequency system master clock, such as 100 MHz. The selection of the system master clock is based on the standard bus frequency of current industrial-grade DSP and FPGA chips. Using this as a high-frequency reference, after processing with the aforementioned dynamic division ratio constant, such as a 32 division, a working clock of approximately 1 MHz to 3 MHz can be generated. This frequency band enters the resonant bandwidth and control response range of the piezoelectric or electromagnetically driven micro two-dimensional galvanometer, ensuring both the smoothness of the follow-up scan and preventing the mechanical actuator from losing synchronization, thereby generating the lower-frequency working clock required to drive the galvanometer scan.

[0063] Finally, the DSP's trajectory synthesis module synthesizes the aforementioned generated reference bias voltage reference value and the high-frequency division ratio constant in a polynomial function superposition manner. In this embodiment, polynomial function superposition refers to a parameterized trajectory generation process, wherein the amplitude term of the trajectory function is determined by the reference bias voltage reference value, and the time frequency term is determined by the working clock derived from the high-frequency division ratio constant.

[0064] Specifically, the reference bias voltage value defines the scan trajectory, such as the maximum displacement in the roll and pitch dimensions as shown in the Lissajous figure, while the high-frequency division ratio constant determines the scan frequency of the trajectory. The DSP combines these parameters into a structured data block containing information such as scan mode, amplitude, and frequency, which serves as a customized physical drive parameter specifically for constraining the roll and pitch displacement steps of the hardware, and outputs it to the downstream actuators.

[0065] Customized physical drive parameters are data structures containing multiple fields, such as {pattern_id,amplitude_x,amplitude_y,clock_divider_ratio}, which are packaged by the DSP and sent to the downstream hardware execution unit via the serial peripheral interface SPI or parallel bus.

[0066] For example, continuing from example S2, the DSP receives a defect spatiotemporal feature data packet containing hard logic morphology encoded words and spatiotemporal stamps. First, the hard logic morphology encoded words are separated, and then the low-frequency distribution coefficient set {3520, -850} and the high-frequency edge coefficient set {125, -90} are extracted from them. Assume the parameters obtained from system calibration are... V / count, V and the inverse scaling system , =0.

[0067] The DSP performs calculations on the reference bias voltage. V, V. For the high-frequency division ratio constant, first calculate the average absolute value of the high-frequency coefficients. Then calculate The DSP combines these calculation results into a data packet of customized physical drive parameters, the contents of which are {pattern_id:1,amplitude_x:4.30,amplitude_y:-1.04,clock_divider_ratio:32}, where pattern_id=1 indicates that the first preset Lissard pattern is used for scanning.

[0068] The data packet was then sent to the 2D galvanometer drive controller, ready for subsequent physical scanning operations. Simultaneously, the spatiotemporal stamp { extracted from the defect spatiotemporal feature data packet} , The value ms is held in the DSP's dedicated register, awaiting the next step.

[0069] In one embodiment of the present invention, the interception trigger signal generation module is configured to perform the following steps: The high-precision pulse count value hidden within the spatial stamp payload data is used to characterize the origin position at the first appearance of the optical phase boundary of the marker. The internal high-speed register area is called to read the difference in physical distance between the proportionally magnified optical detection trigger center and the physical scanning execution station. The compensation offset pulse constant is calculated and entered by the measurement. The hardware direct adder synthesizer is used to combine and enhance the compensation offset pulse constant to the top of the high-precision pulse count value to obtain the expected absolute delay execution threshold. A closed-loop acquisition channel for pulse sequence of digital level of photoelectric drive of large rotating shaft in production is constructed to stably intercept and capture the latest changes in the real-time motion spindle pulse width count value. The zero-wait direct combination gate array is used to seamlessly connect and calibrate the absolute delay execution threshold with the real-time motion spindle pulse width count value. When the value is in the nanosecond-level clock slot corresponding to the matching bit, the rising edge of the level is captured by hardware edge detection logic and the interception trigger signal formed by the drive signal jump is released.

[0070] Specifically, after the physical driving parameter synthesis module generates customized physical driving parameters, the fourth implementation step of this invention is initiated in parallel to monitor and calculate the precise timing for triggering subsequent physical measurements. This step is executed by the aforementioned digital signal processor (DSP) or a closely cooperating FPGA logic unit. The DSP constructs a pulse sequence acquisition closed-loop channel through an input capture or quadrature encoded pulse (QEP) interface. This pulse sequence acquisition closed-loop channel is a hardware-implemented interface capable of decoding the A / B phase quadrature signals output by the encoder in real time and internally maintaining a counter that increments or decrements with the movement of the thin film.

[0071] The physical input of this channel connects to the spindle encoder installed at the downstream physical thickness measurement station. This physical thickness measurement station is the physical location downstream where the confocal spectral measurement probe assembly is installed, and there is a fixed physical distance between this station and the upstream schlieren illumination sensing area. Through this channel, the DSP can continuously read and update the value of an internal register at a microsecond-level refresh rate. This value is the real-time motion spindle pulse width count value at the current position of the physical thickness measurement station. Meanwhile, the DSP parses and extracts from its internal registers the high-precision pulse count value within the spatiotemporal stamp captured in the defect feature encoding and localization module and separated in the physical drive parameter synthesis module. This value represents the spatial location encoding of the defect when it is detected by the upstream schlieren illumination system.

[0072] The DSP reads a pre-calibrated and stored compensation offset pulse constant from its non-volatile memory. This constant is equivalent to the pulse count value corresponding to the fixed spatial physical distance between the optical centerline of the upstream schlieren illumination sensing area and the measurement center point of the downstream physical thickness measurement station on the thin film conveyor line. The DSP's arithmetic logic unit adds the high-precision pulse count value to the compensation offset pulse constant to construct the absolute delay execution threshold. To illustrate more clearly, the absolute delay execution threshold constructed in this step... The calculation follows the formula:

[0073] In the formula, This is a high-precision pulse count value extracted from a spatiotemporal stamp; To compensate for the offset pulse constant. These are key parameters that are calibrated once during the system installation and debugging phase, and their values ​​are obtained by measuring the physical distance between the upstream sensing point and the downstream measuring point. And divide by the unit pulse travel distance determined by the diameter of the spindle encoder and the drive roller. To determine, that is , representing a dynamically calculated target pulse count value.

[0074] This absolute delay execution threshold provides a unique spatial coordinate for each detected defect event, requiring precise execution of the operation at the downstream station. This calculated absolute delay execution threshold, along with the real-time updated real-time motion spindle pulse width count, are continuously fed as two parallel input data streams into a hardwired digital comparator within the DSP.

[0075] Among them, the hardwired digital comparator is a comparison module implemented by hardware logic circuits inside the DSP or FPGA without relying on software instructions. Its comparison speed can reach the system master clock frequency, ensuring zero-latency triggering.

[0076] This digital comparator performs a hardware-level high-low level comparison operation on the bits of two input data within each clock cycle. Within a nanosecond-level clock slot where the bits of the real-time motion spindle pulse width count value and the bits of the absolute delay execution threshold are matched, the comparator outputs a high-level logic signal. The rising edge of this state is captured by a hardware edge detector, which converts it into a narrow-pulse-width high-level logic pulse signal. This signal serves as the nanosecond-level intercept trigger signal, used to instantaneously activate the physical actuator of the in-situ 3D topography acquisition module. This nanosecond-level intercept trigger signal is a standard TTL or LVCMOS level signal, and its effective high-level duration is controlled within one clock cycle, for example, 10 ns under a 100 MHz system clock, to ensure high-precision synchronization characteristics for driving the power switching transistors.

[0077] For example, continuing from the example in S3, the DSP has the customized physical drive parameters ready and maintains a high-precision pulse count value separated from the spatiotemporal stamp. Assuming the system calibration determines the physical distance between the upstream schlieren detection point and the downstream confocal measurement point. The distance traveled per unit pulse of the encoder on the production line is 1.5 m. It is 0.01 mm / pulse.

[0078] Calculate the compensated offset pulse constant One pulse. This value is stored as a fixed system parameter. Subsequently, the DSP calculates the absolute delay execution threshold. The value 12495678 was loaded into one input port of the hardwired digital comparator.

[0079] Meanwhile, the DSP continuously reads the real-time motion spindle pulse width count value generated by the downstream station encoder through its QEP interface. As the film moves, The value keeps increasing. When When the value changes from 12495677 to 12495678, the hard-wired digital comparator detects that the values ​​at the two input ports are equal. At this nanosecond-level instant, the comparator outputs a valid high-level pulse, which is the nanosecond-period intercept trigger signal. This signal is routed to the enable pin of the next stage power drive circuit, indicating that the target defect has reached a position where physical measurements can be performed.

[0080] In one embodiment of the present invention, the in-situ three-dimensional topography acquisition module is used to perform the following operations: In conjunction with in-situ scanning detection, the confocal dispersive sensor probe is simultaneously activated to continuously acquire high-frequency dual-band abnormal spectral feedback data streams of polymer films that are irradiated through the membrane and subsequently reflected and peeled off by the internal structure of the polymer film. The combination of reflection wavelength characteristic values ​​is extracted from the dual-band abnormal spectral feedback data stream, which consists of the first peak center wavelength corresponding to the reflection of the upper surface of the polymer film and the second peak center wavelength corresponding to the reflection of the lower surface. The formula conversion of the offline calibration reference constant coefficient is called, and the absolute difference between the center wavelength of the first peak and the center wavelength of the second peak is compared and calculated. In this way, the absolute thickness difference information of a single detection micro point is decoded. All the absolute thickness difference information captured and collected in the point cloud spatial convergence mapping and integrated coordinate tracking grid are interwoven to construct a three-dimensional physical microscopic morphology data.

[0081] Specifically, in response to the intercept trigger signal with a nanosecond-level period output by the intercept trigger signal generation module, the fifth step of this method is initiated to perform in-situ scanning detection of the target defect region. The steep rising edge generated by the intercept trigger signal is sent to the gate driver of the power switching amplifier matrix network, which acts as an enable signal to turn on the power semiconductor devices, such as MOSFETs or IGBTs, in the network. The power switching amplifier matrix network typically consists of two H-bridge driver circuits, used to drive the X-axis and Y-axis electromagnetic coils of the two-dimensional galvanometer, respectively, and can amplify the low-voltage analog signal from the DAC into a high-precision current signal sufficient to drive the galvanometer.

[0082] The power switching transistor amplifier circuit matrix network, after being turned on, connects the waveform generator, which is in standby mode, to the servo displacement electromagnetic coil of the miniature two-dimensional galvanometer actuator. The miniature two-dimensional galvanometer actuator is a high-speed beam deflection device based on MEMS or electromagnetic drive technology, with a response frequency reaching several kilohertz. The waveform generator has received and loaded customized physical drive parameters in the physical drive parameter synthesis module. These customized physical drive parameters are then parsed and reconstructed into a specific set of analog waveform electrical signals, where the reference bias voltage is converted into the peak voltage of the drive signal, and a high-frequency division ratio constant is used to set the frequency of the drive signal.

[0083] The simulated waveform electrical signal group is streamed into the servo displacement electromagnetic coil of the moving miniature two-dimensional galvanometer actuator. Excited by the asymmetric electromagnetic deflection force field control command caused by the input customized physical driving parameters, the internal scanning mirror of the miniature two-dimensional galvanometer actuator, which carries or is optically coupled to a confocal dispersion sensor probe, begins to follow a non-preset irregular Lissajous figure scanning trajectory synthesized in real time by the optical characteristics of the defect itself. The non-preset irregular Lissajous figure scanning trajectory means that its trajectory shape, amplitude, and frequency are not fixed, but are determined in real time by the customized physical driving parameters unique to each defect in the physical driving parameter synthesis module, thereby achieving adaptive scanning coverage of defects of different shapes.

[0084] The scanning trajectory traverses the physical coordinates of the thin film, which are matched with the spatiotemporal stamps, to target suspected defects. Simultaneously with the galvanometer scanning, a confocal dispersive sensor probe receives the spectral signals penetrating the upper and lower surfaces of the polymer film and reflecting them back at a high sampling frequency, such as tens of thousands of times per second.

[0085] like Figure 3 As shown, the confocal dispersion sensor probe utilizes a specially designed dispersive objective lens to disperse a white light source into a continuous rainbow spectrum along the optical axis, with different wavelengths of light focusing at different depth positions. When light shines on the polymer film surface, only the wavelengths at the focal points are received by the confocal aperture and returned to the spectrometer. Distance can be determined by detecting the center wavelength of this returned light. Figure 3 As can be seen from the polarization characteristic map of the confocal dispersive sensor dual-band reflectance spectrum, for each sampling, the spectral analysis unit inside the probe can identify and extract the polarization characteristic peaks in the two reflectance spectrum waveforms, thereby accurately locating their peak reflection center wavelengths, which correspond to the first peak center wavelength of the reflection from the upper surface of the polymer film and the second peak center wavelength of the reflection from the lower surface, respectively. Together, they constitute the wavelength frequency shift characteristic value combination in the dual-band abnormal spectral feedback data stream.

[0086] Specifically, the frequency shift characteristic value of the reflection center wavelength of the spectral peaks in different bands refers to the wavelength value corresponding to the two intensity peaks detected by the spectrometer in a single acquisition, for example, depending on the physical coordinates of the scanning point. changing and .

[0087] Finally, the spectral reconstruction and decoding unit calls the factory offline calibration reference constant coefficient formula conversion and compares the absolute difference between the center wavelength of the first peak and the center wavelength of the second peak to decode the absolute thickness difference information of a single probe micropoint, i.e., the physical thickness of that point. The absolute thickness difference information obtained at this time is the film thickness at the scanning point. Physical thickness at the location The calculation method is as follows:

[0088] in, The standard offline calibration reference constant coefficients are obtained through calibration using standard thickness samples. Subsequently, the system captures all absolute thickness difference information collected within a point cloud spatial convergence mapping and integrated coordinate tracking grid, interweaving these values ​​to construct a three-dimensional physical microscopic topographic data for further analysis. This physical microscopic topographic data is a two-dimensional array or point cloud set, and its data structure is as follows: ,in It is the position of the scanning point in the galvanometer coordinate system. This is the thickness value at that point. This represents the total number of sampling points.

[0089] For example, continuing from example S4, when the intercept trigger signal with a nanosecond cycle arrives at the power drive board, the prepared customized physical drive parameters {pattern_id:1,amplitude_x:4.30,amplitude_y:-1.04,clock_divider_ratio:32} are converted into analog signals. The peak voltage of the drive signal on the X-axis is 4.30 V, the peak voltage on the Y-axis is -1.04 V, and the scanning clock frequency is 3.125 MHz obtained by dividing the system master clock from 100 MHz by 32. These signals drive the two-dimensional galvanometer, causing the light spot of its mounted confocal dispersion sensor probe to begin mapping on the thin film surface.

[0090] At a certain point on the scanning trajectory The probe's spectrometer receives two beams of reflected light and analyzes them to obtain the peak wavelength from the upper surface of the thin film. nm, peak wavelength from the lower surface nm. Assuming the sensor linearity coefficients are pre-calibrated by the system. The value is 0.1 µm / nm. Therefore, the thickness at that point is calculated using decoding. .

[0091] As the scan progresses, the light spot moves to the center of the defect. At that time, it was measured nm, nm, thickness calculated This indicates the presence of a layer with a thickness of 5.0. The gel points. Throughout the scanning cycle, the system acquired, for example... These are the thickness data points. Together, they form a high-resolution matrix, namely physical microscopic three-dimensional topographic data, which is then transmitted to the decision fusion workstation of the cross-modal fusion and source tracing diagnosis module for final analysis.

[0092] In one embodiment of the present invention, the cross-modal fusion and source tracing diagnostic module is used to perform the following operations: The hard logic morphology encoding is passed into the multi-layer feedforward convolutional neural network processing layer to activate and iterate, deeply mining the abnormal optical flow abrupt change phenomenon characteristics hidden in the gaps of high-dimensional structure. The fusion processing obtains the relative grayscale mapping fingerprint database node data that reflects the variation phenomenon generated by the damage of local microstructure. The optical morphology relative grayscale mapping fingerprint database node data is internally segmented and retained to specifically point to the texture extreme value peak vector that reflects the high density distribution of frequency domain energy and the steep change of tilt angle configuration. At the same time, the physical micro three-dimensional morphology data is extracted and quantified to reflect the spatial thickness anomaly benchmark comparison reference vector composed of the difference of the maximum steep upward gradient of the undulating height surface. The calibration and correction of the texture extreme peak vector and the comparison of the spatial thickness anomaly reference vector are forced to be registered and mapped within the same virtual multifaceted analytical coordinate system. The evolution of the two vectors is observed, and their pointing angles are aligned and synchronized. The deviation of the trajectory changes is measured by algebraic embedding of Pearson coefficients. The measurement model is used to calculate the mapping, quantitatively presenting the density of the spatial distribution variation and the level of linear fitting overlap compatibility between the two vectors. The output is an optical microscopic thickness compatibility correlation coefficient value that characterizes the internal substantial consistency confidence intensity.

[0093] Specifically, after the in-situ 3D topography acquisition module generates physical microscopic 3D topography data, the final step of this method is executed. This step aims to fuse cross-modal data and perform causal attribution discrimination. This process is performed by a tensor analysis processor located at the core of the decision fusion workstation, equipped with a high-performance graphics processing unit (GPU) or a dedicated tensor processing unit (TPU).

[0094] The tensor analysis processor invokes a pre-trained deep convolutional neural network model to perform deep feature extraction on the hard logic morphology codes associated with the current defect event generated in the defect feature encoding and localization module. The hard logic morphology codes are parsed and restored to a frequency domain coefficient matrix by the model, serving as the network's input layer. Through multiple convolutional and pooling operations, the network deeply extracts and amplifies the high-dimensional anomalous structural coefficient combinations exhibited in the frequency domain space by optical abrupt changes. The network's final output is a high-dimensional feature vector, constructed as a relative grayscale fingerprint database node data describing the optical morphology of the thin film under microstructural stress or structural defects. The optical morphology relative grayscale fingerprint database node data is a high-dimensional vector, for example, with dimensions of 1024, obtained by feature extraction of the DCT coefficient matrix using a pre-trained model such as ResNet or EfficientNet.

[0095] The tensor analysis processor aligns and calibrates two datasets from different sources within a unified Euclidean coordinate system. The first is a segmentation of the optical morphology relative grayscale mapping fingerprint database node data generated by the aforementioned module, specifically retaining texture extremum peak vectors that reflect the abrupt changes in tilt angle configuration reflecting the highly dense distribution of frequency domain energy. These texture extremum peak vectors are subsets of some dimensions of the aforementioned high-dimensional vectors, representing the direction where frequency domain energy is most concentrated and texture changes drastically. Secondly, a spatial thickness anomaly benchmark comparison reference vector is simultaneously extracted from the physical microscopic three-dimensional topography data, which quantitatively reflects the difference in the maximum steepness of the undulating height surface, forming a spatial thickness anomaly benchmark comparison reference vector. This spatial thickness anomaly benchmark comparison reference vector is derived from the three-dimensional features representing the evolution of the maximum gradient difference at physical concavity and convexity positions, extracted from the digital elevation model (DEM) generated by surface fitting or Delaunay triangulation of the point cloud data obtained by the in-situ three-dimensional topography acquisition module.

[0096] After extracting the two sets of vectors, the processor performs calibration and correction, forcibly registering and mapping the texture extreme peak vector and the spatial thickness anomaly benchmark comparison reference vector within the same virtual multifaceted analytical coordinate system scale. This allows for the observation of the amplitude trajectory of their directional angle alignment and synchronous path deviation. Subsequently, an algebraic embedding Pearson coefficient is used to calculate the metric model and estimate the mapping, quantitatively presenting the spatial distribution variation density and linear fitting overlap compatibility level of the two vectors. The output is an optical microscopic thickness compatibility correlation coefficient value representing the internal substantial consistency confidence intensity.

[0097] The Pearson correlation algorithm is a statistical method used to measure the degree of linear correlation between two variables. The calculated result, the linear fit degree in the correlation space, ranges from -1 to +1. A value close to +1 indicates a positive correlation between optical anomalies and physical convex height, close to -1 indicates a positive correlation with physical concave height, and close to 0 indicates no significant linear relationship. This coefficient quantifies the synchronicity and consistency between the degree of anomaly in the optical characterization and the changes in thickness measured physically.

[0098] The tensor analysis processor compares and retrieves the calculated optical micro-thickness compatibility correlation coefficient values ​​and their derived evolution matrices over time or batches with an offline-constructed prior database of production parameter failure models. This prior database is a knowledge base built through numerous offline experiments. It maps various known process parameter anomalies, such as impurities in raw material A or uneven cooling in region B, to the optical fingerprint-physical morphology dual characteristics of defects in the final product, forming different clusters in a multi-dimensional feature space. This database stores a large number of clustered scatter plots representing known process defects, such as raw material contamination and extrusion temperature fluctuations, calibrated through experiments or historical data.

[0099] The processor determines the degree of agreement between the characteristics of the current defect and the Mahalanobis distance or other similarity measures of each group in the database. By setting a preset agreement threshold, the processor can identify and separate real internal hard free gel points whose optical features are highly correlated with changes in physical thickness, and distinguish them from surface stress reflection marks that have only strong optical features but no significant changes in physical thickness. This agreement threshold, set at 85%, is determined based on offline large-sample receiver operating characteristic curve analysis. By comparing real defects and optical artifacts verified by extensive manual re-inspection in a priori database, it was found that when the agreement calculated by the Pearson correlation coefficient and Mahalanobis distance is set in the range of 80% to 90%, the system can effectively filter out 99% of optical false alarms caused by minor airflow fluctuations or harmless surface scratches on the film surface, while keeping the false negative rate of real structural damage defects below the tolerance limit.

[0100] Once a genuine internal defect is identified and its characteristics closely match a high-risk process failure mode in the database, the system will issue a targeted alarm and output a final-state defect tracing warning that includes the defect type, possible causes, and suggested process adjustment measures, thereby completing the entire detection and diagnosis process.

[0101] This final state defect tracing warning is a structured information package that includes not only the coordinates and severity level of the defect, but may also include specific diagnostic information such as: Warning: C-type gel point detected, with 95% consistency with the contamination characteristics of raw material batch P-2, recommending immediate shutdown and replacement of raw material, etc.

[0102] For example, continuing from example S5, the decision fusion workstation receives the hard logic morphological code {3520,-850,125,-90} associated with the defect, as well as physical microscopic three-dimensional morphological data. The thickness matrix has a center thickness of 5.0 μm and a perimeter thickness of 2.0 μm. Hard logic morphological encoded words are fed into a feature extraction model to generate a 1024-dimensional optical morphology relative grayscale mapping fingerprint database node data. Simultaneously, the physical microscopic three-dimensional morphology data is reconstructed in three dimensions, forming a mountain-like contour with a central convex protrusion.

[0103] Calculations showed that the angle between the direction of the most drastic texture change indicated by the optical feature vector and the direction of the largest gradient on the physical contour was less than 5 degrees. Then, the Pearson correlation algorithm was used to calculate the consistency of their changing trends, yielding an optical micro-thickness compatibility correlation coefficient of +0.92. This positively correlated value objectively indicates that the initially detected optical anomaly was indeed caused by a real physical protrusion.

[0104] Next, the feature vector of this defect, including the optical fingerprint and correlation coefficient value, was compared with the prior database of the production parameter failure model. It was found that the position of this feature vector in the feature space was closest to the center of the cluster of scatter plots in the database labeled as hard gel points caused by catalyst residue in BOPP film raw material T-300S, with a consistency of 98%. The set consistency threshold was 85%. Since 98% is greater than 85%, the system determined that this defect was a real, high-risk internal hard gel point. Therefore, a final-state defect tracing warning popped up on the interface in the central monitoring room, displaying: Critical Warning! A 5.0 μm Class A gel point was found at coordinates (12345678). Cause tracing: Highly consistent with the catalyst residue defect model. Recommendation: Immediately sample the current batch of raw materials and check the cleaning procedures of the upstream reactor.

[0105] At the same time, the defect is automatically recorded in the batch quality report, and the production line cutting or marking mechanism is triggered to physically isolate the section of film.

[0106] See appendix Figure 2 This invention proposes a method for detecting the thickness quality of a capacitor polymer film, comprising the following steps: S1. Obtain the internal refractive micro-transmitted light flow through the polymer film and convert it into a continuous grayscale sequence. Calculate the total value of the light intensity gradient energy corresponding to the continuous grayscale sequence. When the total value of the light intensity gradient energy exceeds the preset energy threshold boundary, generate a defect event judgment command. S2. The defect event judgment instruction cuts out the solidified abnormal image frame from the continuous grayscale sequence, performs frequency domain mapping and splicing on the solidified abnormal image frame to generate hard logic morphology code word, combines the preset high-precision pulse count value and global timestamp to construct a spatiotemporal stamp, and encapsulates the hard logic morphology code word and spatiotemporal stamp to generate defect spatiotemporal feature data packet. S3. Unpack the defective spatiotemporal feature data packet to separate the hard logic morphology code word and spatiotemporal stamp, and extract the frequency domain coefficients in the hard logic morphology code word to map and synthesize customized physical driving parameters; S4. Obtain the real-time motion spindle pulse width count value, extract the high-precision pulse count value inside the time stamp and superimpose the preset compensation offset pulse constant to generate an absolute delay execution threshold, and compare the absolute delay execution threshold with the real-time motion spindle pulse width count value to generate an interception trigger signal. S5 responds to the interception trigger signal and controls the probe to perform in-situ scanning detection according to the customized physical driving parameters. It receives the reflection spectrum decoding to calculate the absolute thickness difference information and compiles it into physical microscopic three-dimensional morphology data. S6. Extract the structural features of the hard logic morphological encoded words to establish fingerprint database node data. Combine the fingerprint database node data with the physical microscopic three-dimensional morphology data to calculate the optical microscopic thickness compatibility correlation coefficient value. Substitute the optical microscopic thickness compatibility correlation coefficient value into the preset offline production parameter failure model prior database for comparison and judgment, and output the final state defect tracing warning.

[0107] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0108] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A capacitor polymer film thickness quality detection system, characterized in that, include: The real-time optical anomaly detection module acquires the internal refractive micro-transmission light flow through the polymer film and converts it into a continuous grayscale sequence. It calculates the total value of the light intensity gradient energy corresponding to the continuous grayscale sequence and generates a defect event judgment command when the total value of the light intensity gradient energy exceeds the preset energy threshold boundary. The defect feature encoding and localization module responds to defect event judgment instructions by cutting out solidified abnormal image frames from a continuous grayscale sequence, performing frequency domain mapping and splicing on the solidified abnormal image frames to generate hard logic morphology encoding words, combining preset high-precision pulse count values ​​and global timestamps to construct a spatiotemporal stamp, and encapsulating the hard logic morphology encoding words and spatiotemporal stamps to generate a defect spatiotemporal feature data packet. The physical driving parameter synthesis module unpacks the defect spatiotemporal feature data packet, separates the hard logic morphology code word and spatiotemporal stamp, and extracts the frequency domain coefficients in the hard logic morphology code word to synthesize customized physical driving parameters. The intercept trigger signal generation module acquires the real-time motion spindle pulse width count value, extracts the high-precision pulse count value inside the time stamp and superimposes it with the preset compensation offset pulse constant to generate an absolute delay execution threshold, and compares the absolute delay execution threshold with the real-time motion spindle pulse width count value to generate an intercept trigger signal. The in-situ three-dimensional topography acquisition module responds to the interception trigger signal and controls the probe to perform in-situ scanning detection according to the customized physical driving parameters. It receives the reflection spectrum decoding to calculate the absolute thickness difference information and compiles it into physical microscopic three-dimensional topography data. The cross-modal fusion and source tracing diagnosis module extracts the structural features of hard logic morphological encoded words to establish fingerprint database node data. It combines the fingerprint database node data with physical microscopic three-dimensional morphology data to calculate the optical microscopic thickness compatibility correlation coefficient value. The optical microscopic thickness compatibility correlation coefficient value is substituted into the preset offline production parameter failure model prior database for comparison and judgment, and the final state defect source tracing warning is output.

2. The capacitor polymer film thickness quality detection system according to claim 1, characterized in that, Obtaining the internal refractive micro-transmission light flow through the polymer film and converting it into a continuous grayscale sequence includes the following steps: The driving schlieren illumination optical array projects a pre-defined structured illumination field onto the traveling polymer film; Simultaneously activate a high-speed linear array camera to capture the internally refracted microscopic transmitted light flow that is deflected through the polymer film; The spatial variation modes of the internal refractive micro-transmitted light flow are converted into digital light intensity amplitudes, and a continuous grayscale sequence reflecting the physical parameters of the first derivative of the polymer film thickness is constructed using the digital light intensity amplitudes.

3. The capacitor polymer film thickness quality detection system according to claim 1, characterized in that, When the total energy value of the light intensity gradient exceeds a preset energy threshold boundary, a defect event determination instruction is generated, including the following steps: The continuous grayscale sequence is scanned one by one using a sliding detection window with a pre-set arrangement pattern size, and the analysis pixel matrix space is extracted and separated. Calculate the approximate gray-level difference gradient of all pixels across the horizontal and vertical ends within the sliding detection window, calculate the square of each approximate gray-level difference gradient individually, and sum them up across the entire sliding detection window to obtain a continuous and smooth total value of light intensity gradient energy. The total value of light intensity gradient energy obtained in real time is continuously sent to the hardware digital comparator to match and check the preset energy threshold boundary. When the total value of light intensity gradient energy is detected to be greater than the energy threshold boundary, the pin feedback outputs a level flip signal to form and transmit the defect event judgment instruction.

4. The capacitor polymer film thickness quality detection system according to claim 1, characterized in that, The process of performing frequency domain mapping and splicing on solidified abnormal image frames to generate hard logic morphological codes includes the following steps: The solidified abnormal image frame is imported into a proprietary two-dimensional discrete cosine transform calculation core array to perform a two-dimensional spatial to frequency domain transformation, generating a transform frequency domain matrix that is laid out by energy coefficients of different frequency bands. Extract the set of low-frequency distribution coefficients that represent the width of the macroscopic geometric distribution of suspected morphological defects inside the transform frequency domain matrix, and simultaneously extract the set of high-frequency edge coefficients that express the sharpness attribute of the abrupt change of the contour edge in the high-frequency distribution region at the very end of the transform frequency domain matrix. A digital fixed-point alignment bit-width calling logic cross-shift dynamic calculation combination splicing merge low-frequency distribution coefficient set and high-frequency edge coefficient set is established, integrated, compressed and packaged into a binary mask block of fixed encapsulation length and distributed externally as a hard logic morphological encoding word.

5. The capacitor polymer film thickness quality detection system according to claim 1, characterized in that, Extracting frequency domain coefficients from hard logic morphological encoded words and mapping them to synthesize customized physical driving parameters includes the following steps: The low-frequency distribution coefficient set, which records the macroscopic surface expansion contour features of defects and is encapsulated inside the hard logic morphological encoded word, is extracted and transformed by linear translation, superposition, and scaling, and then assigned to the reference bias voltage reference value of the galvanometer digital-analog core power supply simulation reference. Extract the high-frequency edge coefficient set that represents the sharp intensity of micro-abrupt changes at the boundary of the quantization expression within the hard logic morphology encoded word. Normalize the high-frequency edge coefficient set to obtain the average absolute floating-point constant and perform inverse linear scaling to convert it into the high-frequency division ratio constant that the internal clock flow of the driving system depends on. The high-frequency division ratio constant value is inversely proportional to the coefficient value in the high-frequency edge coefficient set. By establishing a mathematical hierarchy-related polynomial superposition and overall allocation program, the reference bias voltage base value is allocated and limited to the maximum coverage limit of the main sampling window displacement. The high-frequency division ratio constant is changed to the physical position update refresh frequency beat parameter. The integrated mixed configuration guides the subsequent probe physical follow-up execution of exclusive customized physical drive parameters.

6. The capacitor polymer film thickness quality detection system according to claim 1, characterized in that, The process of generating an intercept trigger signal by comparing the absolute delay execution threshold with the real-time motion spindle pulse width count value includes the following steps: The high-precision pulse count value hidden within the spatial stamp payload data is used to characterize the origin position at the first appearance of the optical phase boundary of the marker. The internal high-speed register area is called to read the difference in physical distance between the proportionally magnified optical detection trigger center and the physical scanning execution station. The compensation offset pulse constant is calculated and entered by the measurement. The hardware direct adder synthesizer is used to combine and enhance the compensation offset pulse constant to the top of the high-precision pulse count value to obtain the expected absolute delay execution threshold. A closed-loop acquisition channel for pulse sequence of digital level of photoelectric drive of large rotating shaft in production is constructed to stably intercept and capture the latest changes in the real-time motion spindle pulse width count value. The zero-wait direct combination gate array is used to seamlessly connect and calibrate the absolute delay execution threshold with the real-time motion spindle pulse width count value. When the value is in the nanosecond-level clock slot corresponding to the matching bit, the rising edge of the level is captured by hardware edge detection logic and the interception trigger signal formed by the drive signal jump is released.

7. The capacitor polymer film thickness quality detection system according to claim 1, characterized in that, Responding to the interception trigger signal and controlling the probe to perform in-situ scanning detection according to customized physical drive parameters includes the following steps: The power switching transistor amplifier circuit matrix network is awakened and activated by the power jump event that occurs at the front end of the interception trigger signal, which is driven by the bottom execution element that is waiting to be taken over. The customized physical drive parameters are unpacked and reassembled to generate a set of analog waveform electrical signals with smooth high and low undulations and simulated control trajectories. These signals are then streamed into the miniature two-dimensional galvanometer displacement servo electromagnetic coil that supports the rocker arm and changes its direction. By injecting a group of analog waveform electrical signals into the displacement servo electromagnetic coil of the miniature two-dimensional galvanometer, an asymmetric electromagnetic deflection force field with uneven distribution and rapid jumps is generated. This force commands the miniature two-dimensional galvanometer actuator to drive the confocal dispersion sensor probe mounted on the side of the miniature two-dimensional galvanometer actuator to repeatedly sweep across the set area following an irregular Lissajous figure scanning trajectory generated by the real-time non-periodic changes of customized physical driving parameters. This achieves a comprehensive and full-range inspection of the source of the anomaly, completing the required in-situ scanning detection.

8. The capacitor polymer film thickness quality detection system according to claim 1, characterized in that, The data is compiled into physical microscopic three-dimensional topographic data, including the following steps: In conjunction with in-situ scanning detection, the confocal dispersive sensor probe is simultaneously activated to continuously acquire high-frequency dual-band abnormal spectral feedback data streams of polymer films that are irradiated through the membrane and subsequently reflected and peeled off by the internal structure of the polymer film. The combination of reflection wavelength characteristic values ​​is extracted from the dual-band abnormal spectral feedback data stream, which consists of the first peak center wavelength corresponding to the reflection of the upper surface of the polymer film and the second peak center wavelength corresponding to the reflection of the lower surface. The formula conversion of the offline calibration reference constant coefficient is called, and the absolute difference between the center wavelength of the first peak and the center wavelength of the second peak is compared and calculated. In this way, the absolute thickness difference information of a single detection micro point is decoded. All the absolute thickness difference information captured and collected in the point cloud spatial convergence mapping and integrated coordinate tracking grid are interwoven to construct a three-dimensional physical microscopic morphology data.

9. The capacitor polymer film thickness quality detection system according to claim 1, characterized in that, The optical micro-thickness compatibility correlation coefficient is calculated by combining fingerprint database node data with physical micro-three-dimensional morphology data, including the following steps: The hard logic morphology encoding is passed into the multi-layer feedforward convolutional neural network processing layer to activate and iterate, deeply mining the abnormal optical flow abrupt change phenomenon characteristics hidden in the gaps of high-dimensional structure. The fusion processing obtains the relative grayscale mapping fingerprint database node data that reflects the variation phenomenon generated by the damage of local microstructure. The optical morphology relative grayscale mapping fingerprint database node data is internally segmented and retained to specifically point to the texture extreme value peak vector that reflects the high density distribution of frequency domain energy and the steep change of tilt angle configuration. At the same time, the physical micro three-dimensional morphology data is extracted and quantified to reflect the spatial thickness anomaly benchmark comparison reference vector composed of the difference of the maximum steep upward gradient of the undulating height surface. The calibration and correction of the texture extreme peak vector and the comparison of the spatial thickness anomaly reference vector are forced to be registered and mapped within the same virtual multifaceted analytical coordinate system. The evolution of the two vectors is observed, and their pointing angles are aligned and synchronized. The deviation of the trajectory changes is measured by algebraic embedding of Pearson coefficients. The measurement model is used to calculate the mapping, quantitatively presenting the density of the spatial distribution variation and the level of linear fitting overlap compatibility between the two vectors. The output is an optical microscopic thickness compatibility correlation coefficient value that characterizes the internal substantial consistency confidence intensity.

10. A method for detecting the thickness quality of a capacitor polymer film, characterized in that, Includes the following steps: S1. Obtain the internal refractive micro-transmitted light flow through the polymer film and convert it into a continuous grayscale sequence. Calculate the total value of the light intensity gradient energy corresponding to the continuous grayscale sequence. When the total value of the light intensity gradient energy exceeds the preset energy threshold boundary, generate a defect event judgment command. S2. The defect event judgment instruction cuts out the solidified abnormal image frame from the continuous grayscale sequence, performs frequency domain mapping and splicing on the solidified abnormal image frame to generate hard logic morphology code word, combines the preset high-precision pulse count value and global timestamp to construct a spatiotemporal stamp, and encapsulates the hard logic morphology code word and spatiotemporal stamp to generate defect spatiotemporal feature data packet. S3. Unpack the defective spatiotemporal feature data packet to separate the hard logic morphology code word and spatiotemporal stamp, and extract the frequency domain coefficients in the hard logic morphology code word to map and synthesize customized physical driving parameters; S4. Obtain the real-time motion spindle pulse width count value, extract the high-precision pulse count value inside the time stamp and superimpose the preset compensation offset pulse constant to generate an absolute delay execution threshold, and compare the absolute delay execution threshold with the real-time motion spindle pulse width count value to generate an interception trigger signal. S5 responds to the interception trigger signal and controls the probe to perform in-situ scanning detection according to the customized physical driving parameters. It receives the reflection spectrum decoding to calculate the absolute thickness difference information and compiles it into physical microscopic three-dimensional morphology data. S6. Extract the structural features of the hard logic morphological encoded words to establish fingerprint database node data. Combine the fingerprint database node data with the physical microscopic three-dimensional morphology data to calculate the optical microscopic thickness compatibility correlation coefficient value. Substitute the optical microscopic thickness compatibility correlation coefficient value into the preset offline production parameter failure model prior database for comparison and judgment, and output the final state defect tracing warning.

Images